💻 computer science

FLUX: Accelerating Cross-Embodiment Generative Navigation Policies via Rectified Flow and Static-to-Dynamic Learning

This paper introduces FLUX, a unified flow-based navigation policy that leverages rectified flow for efficient inference and a static-to-dynamic curriculum to achieve state-of-the-art performance across six navigation tasks while demonstrating zero-shot sim-to-real transfer across diverse robotic embodiments.

Original authors: Zeying Gong, Yangyi Zhong, Yiyi Ding, Tianshuai Hu, Guoyang Zhao, Lingdong Kong, Rong Li, Jiadi You, Junwei Liang

Published 2026-03-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zeying Gong, Yangyi Zhong, Yiyi Ding, Tianshuai Hu, Guoyang Zhao, Lingdong Kong, Rong Li, Jiadi You, Junwei Liang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to navigate a busy city. You want it to be able to do two very different things:

The Quiet Library: Walk straight to a specific book on a shelf without bumping into anything (Static Navigation).
The Busy Subway: Weave through a rushing crowd of people, dodging strangers while still trying to get to your destination (Dynamic Social Navigation).

Most robots today are like students who only studied for one of these exams. If you teach them to walk in a library, they panic in a crowd. If you teach them to dodge people, they get lost in a quiet room.

The paper introduces FLUX, a new "super-robot brain" that is great at both, and it does so using some clever tricks. Here is how it works, explained simply:

1. The Problem: The Robot's "Stutter"

Older robot brains (called Diffusion Models) work like a person trying to draw a picture by starting with a messy scribble and slowly erasing the noise until the image appears. They have to do this over and over again, step-by-step, to figure out where to go.

The Analogy: It's like trying to walk to a store by taking 50 tiny, hesitant steps, checking your map, stepping back, and checking again. It's accurate, but it's slow.

2. The Solution: The "Straight Line" Shortcut (Rectified Flow)

The authors realized they could teach the robot to draw a straight line from "Confused" to "Destination" instead of a messy scribble.

The Analogy: Instead of the hesitant 50 steps, FLUX is like a high-speed train on a straight track. It cuts out all the wiggles and noise.
The Result: This makes the robot 47% faster at making decisions than previous methods. It can think and move almost instantly, which is crucial when people are running around you.

3. The Training Method: "Static-to-Dynamic" School

How did they teach this robot to be so good? They used a special two-step curriculum, like a martial arts master training a student.

Step 1: The Geometry Class (Static Learning)
First, they teach the robot in a quiet, empty room. They show it how to walk in straight lines and avoid walls.
- The Metaphor: This is like learning to drive in an empty parking lot. The robot learns the basic "rules of the road" and how to get from A to B efficiently.
Step 2: The Rush Hour Class (Dynamic Learning)
Next, they throw the robot into a simulation of a busy crowd. But here is the magic: because the robot already knows the basics from Step 1, it doesn't panic. It learns to "dance" around people.
- The Metaphor: This is like taking your driving skills to a busy highway. Because you already know how to steer, you can now focus entirely on dodging other cars.
- The Surprise: The paper found that training in the "busy crowd" actually made the robot better at the "empty parking lot" tasks too! It learned to be more careful and recover from mistakes, making it a safer driver everywhere.

4. The Benchmark: "DynBench"

To prove this works, the team built a new testing ground called DynBench.

The Analogy: Imagine a video game level where the NPCs (non-player characters) aren't just walking in circles or following simple scripts. They are "realistic." They stop to look at phones, change direction suddenly, and bump into each other naturally.
This new test ground forced the robot to prove it could handle real-world chaos, not just a fake, easy version of it.

5. The Real-World Test: One Brain, Three Bodies

The coolest part? They took this single "brain" (the software) and put it on three completely different types of robots:

A Wheeled Robot: Smooth and stable (like a Roomba).
A Quadruped Robot: A dog-like robot that bounces and shakes the camera (like a Boston Dynamics Spot).
A Humanoid Robot: A robot that walks on two legs and wobbles (like a human).

The Result: The robot didn't need to be re-taught or "fine-tuned" for each body. It just worked.

The Analogy: It's like giving a human the same driving instructions whether they are driving a sedan, a pickup truck, or a motorcycle. The core logic of "don't hit the other car" remains the same, regardless of the vehicle.

Summary

FLUX is a new navigation system that:

Thinks Faster: Uses a "straight line" math trick to make decisions instantly.
Learns Better: Starts by learning the basics in a quiet room, then masters the chaos of a crowd, which makes it safer everywhere.
Works Everywhere: Can control a wheel, a dog-leg, or a human-leg robot without needing a new manual for each one.

It's essentially teaching robots to be the ultimate "city commuters" who can handle both a quiet library and a rush-hour subway with equal ease.

1. Problem Statement

Autonomous navigation requires a diverse set of skills, ranging from static goal-reaching to dynamic social traversal in human-populated environments. Current research suffers from fragmentation:

Policy Limitations: Traditional Reinforcement Learning (RL) lacks generalization across tasks; Imitation Learning (IL) struggles with dynamic social scenarios; and existing generative models (e.g., Diffusion) suffer from high inference latency due to iterative denoising.
Evaluation Gaps: Existing benchmarks are either static (ignoring social constraints) or use oversimplified crowd models (e.g., random walks or ORCA planners) that fail to capture realistic human dynamics.
Generalization: There is a lack of unified protocols that evaluate policies across diverse task modalities and heterogeneous robot platforms (wheeled, quadrupedal, humanoid).

The core challenge is to develop a unified navigation policy that is both computationally efficient (real-time capable) and robust to dynamic, multi-modal environments, while enabling zero-shot transfer across different robot embodiments.

2. Methodology: The FLUX Framework

The authors propose FLUX (Flow-based Unified policy for X-platform), a generative navigation policy that employs a static-to-dynamic curriculum and Rectified Flow for efficient inference.

A. Core Architecture: Rectified Flow

Instead of using iterative diffusion models (which require many steps to denoise), FLUX utilizes Rectified Flow (RF) to linearize the probability flow between noise and data.

Mechanism: RF learns a velocity field $v_\theta$ that transports noise $\tau_0$ to expert trajectories $\tau_1$ along a straight line (linear interpolant).
Benefit: This replaces the curved stochastic transport path of diffusion models with straight-line trajectories. Consequently, FLUX achieves high-quality generation with significantly fewer sampling steps (Euler integration), drastically reducing inference latency.
Generative Nature: The policy outputs a distribution of candidate trajectories rather than a single deterministic action, allowing for multi-modal responses essential for deadlock recovery and proactive avoidance.

B. Two-Stage Training Paradigm

FLUX follows a progressive training strategy:

Stage 1: Flow-Based Imitation Learning (Static):
- The model is pre-trained on static expert trajectories using Rectified Flow.
- Goal: Establish strong geometric priors for path planning and goal-reaching.
- Input: Egocentric visual observations (RGB-D) and goal conditions.
Stage 2: GRPO-Based Reinforcement Learning (Dynamic):
- The policy is fine-tuned in dynamic environments using Group Relative Policy Optimization (GRPO).
- Mechanism: The generative backbone produces $M$ candidate trajectories. These serve as a natural "group" for advantage estimation, eliminating the need for a separate value-function critic.
- Reward Design: Includes asymmetric goal progress, proxemic social penalties (based on Hall's zones), and exploration bonuses.
- Goal: Refine the policy to handle social interactions, dynamic obstacles, and recovery behaviors.

3. Key Contributions

FLUX Policy: The first flow-based unified navigation policy that bridges multi-modal trajectory generation with efficient robotic control. It achieves 47% higher inference efficiency than prior flow-based methods and 29% over diffusion-based methods.
DynBench Benchmark: A new benchmark featuring physically valid crowd simulation (using a hierarchical Markov FSM and NavMesh-based planning) that complements existing static protocols. It supports six fundamental tasks: PointNav, ImageNav, Exploration, Dynamic PointNav, Dynamic Exploration, and SocialNav.
Static-to-Dynamic Curriculum: The paper demonstrates that training on static data first, followed by RL refinement in dynamic environments, yields a positive transfer effect. The dynamic training not only improves social navigation but also enhances robustness in static tasks by capturing recovery behaviors.
Cross-Embodiment Zero-Shot Transfer: FLUX demonstrates successful zero-shot sim-to-real transfer across three distinct morphologies (wheeled, quadrupedal, and humanoid) without any platform-specific fine-tuning.

4. Experimental Results

Performance in Simulation (DynBench)

FLUX achieved State-of-the-Art (SOTA) performance across all six tasks:

Static Tasks: Outperformed baselines (including NavDP) in Success Rate (SR) and Success weighted by Path Length (SPL). For example, PointNav SR improved by 3.1% over NavDP.
Dynamic Tasks: Showed superior robustness. In SocialNav, FLUX achieved a 64.0% SR (vs. 59.3% for NavDP) with a significant reduction in collisions (16 vs. 21) and improved social compliance.
Efficiency: With only 6 sampling steps, FLUX maintained high performance while reducing inference time to 216.5ms, compared to 305.2ms for NavDP (at 10 steps).

Real-World Deployment

Platforms: Deployed on Unitree Go2-W (wheeled), Unitree Go2 (quadrupedal), and Unitree G1 (humanoid).
Result: The policy successfully navigated real-world corridors and offices with static obstacles and dynamic pedestrians.
Significance: Despite significant differences in camera height, gait cycles, and kinematics, FLUX transferred zero-shot without fine-tuning, proving the robustness of the unified generative approach.

Ablation Studies

Generative vs. Deterministic: Replacing the generative backbone with deterministic regression caused a significant performance drop, confirming the necessity of modeling action stochasticity.
Rectified Flow vs. Diffusion/CFM: Rectified Flow provided consistent gains over standard Conditional Flow Matching (CFM) and Diffusion models, validating the straight-line transport assumption.
RL Refinement: The GRPO post-training stage was crucial for improving social compliance and dynamic exploration metrics.

5. Significance and Impact

Efficiency Revolution: By linearizing probability flow, FLUX solves the "latency bottleneck" of generative navigation, making real-time, multi-modal decision-making feasible for resource-constrained robots.
Unified Framework: It challenges the fragmented landscape of navigation research by providing a single policy capable of handling static, dynamic, social, and exploratory tasks.
Generalization: The successful zero-shot transfer across heterogeneous robots suggests that learning a unified "flow" of navigation behaviors is more effective than training task-specific or embodiment-specific policies.
Benchmarking: DynBench sets a new standard for evaluating navigation by incorporating physically realistic human dynamics, moving beyond simplified rule-based crowd models.

In conclusion, FLUX represents a significant leap forward in autonomous navigation, combining the efficiency of rectified flow with the robustness of a static-to-dynamic learning curriculum to create a versatile, safe, and efficient navigation policy for the real world.